Publicação
Integrating knowledge of database systems with language models
| Resumo: | Recent advances in Artificial Intelligence have enabled large language model (LLM) agents to query structured data sources, such as SQL databases or APIs, based on natural language prompts. This capability provides a richer context and improves response quality. This work explores a dynamic approach to translating user prompts into data queries that operate, manage, and retrieve information beyond the capabilities of pre-built tools. To fulfill this goal, several well-established approaches have been developed and applied in different contexts, such as Text-2-SQL, API based approaches, semantic parsing, among others. However, many implementations are troubled with security issues, as LLM-based systems are vulnerable to scenarios such as prompt injection, which can disrupt the entire system, exfiltrate restricted data, and allow for many other dangerous possibilities. Because this work aims to design systems intended for real operational use, addressing such security risks is a central concern of the dissertation. This dissertation is conducted in collaboration with OmniumAI, the developers of CibusAI, an artificial intelligence based platform that empowers food science by supporting product reformulations, promoting sustainability, and optimizing various processes. The work focuses on enhancing the capabilities of the CibusAI platform’s AI assistant, giving it the ability to dynamically query stored data. Through the design and implementation of a dynamic query system integrated into CibusAI, this work demonstrates how LLM-powered agents can be extended to interact with structured data sources in an industrial setting. The system was evaluated with Langfuse across multiple datasets and languages, including robustness checks against adversarial attacks. Although the developed tools, for advanced formulation development scenarios, showed limited accuracy and faced challenges such as latency, token constraints, and inadequate tool selection, the evaluation highlighted a high degree of relevance and revealed practical strategies for improving tool documentation, endpoint design, and security hardening. These results underline both the potential and complexity of deploying secure, dynamic LLM query systems in real-world environments, providing lessons and design guidelines that inform future research and industrial adoption. |
|---|---|
| Autores principais: | Silva, Miguel Filipe Cidade da |
| Assunto: | AI LLM RAG NLP Prompt Injection Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Recent advances in Artificial Intelligence have enabled large language model (LLM) agents to query structured data sources, such as SQL databases or APIs, based on natural language prompts. This capability provides a richer context and improves response quality. This work explores a dynamic approach to translating user prompts into data queries that operate, manage, and retrieve information beyond the capabilities of pre-built tools. To fulfill this goal, several well-established approaches have been developed and applied in different contexts, such as Text-2-SQL, API based approaches, semantic parsing, among others. However, many implementations are troubled with security issues, as LLM-based systems are vulnerable to scenarios such as prompt injection, which can disrupt the entire system, exfiltrate restricted data, and allow for many other dangerous possibilities. Because this work aims to design systems intended for real operational use, addressing such security risks is a central concern of the dissertation. This dissertation is conducted in collaboration with OmniumAI, the developers of CibusAI, an artificial intelligence based platform that empowers food science by supporting product reformulations, promoting sustainability, and optimizing various processes. The work focuses on enhancing the capabilities of the CibusAI platform’s AI assistant, giving it the ability to dynamically query stored data. Through the design and implementation of a dynamic query system integrated into CibusAI, this work demonstrates how LLM-powered agents can be extended to interact with structured data sources in an industrial setting. The system was evaluated with Langfuse across multiple datasets and languages, including robustness checks against adversarial attacks. Although the developed tools, for advanced formulation development scenarios, showed limited accuracy and faced challenges such as latency, token constraints, and inadequate tool selection, the evaluation highlighted a high degree of relevance and revealed practical strategies for improving tool documentation, endpoint design, and security hardening. These results underline both the potential and complexity of deploying secure, dynamic LLM query systems in real-world environments, providing lessons and design guidelines that inform future research and industrial adoption. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.